DE69031523T2 - Digital image display device - Google Patents

Description

The present invention relates to a digital image display device for displaying an image by converting a pixel value of digital image data into a brightness by means of a display window, in which a window level and a window width can be suitably set as parameters relating to the conversion of the pixel value into the brightness, that is, relating to the display window, depending on characteristics of the digital image data.

An example of an image display device of such an embodiment is a device for displaying a medical image such as a CT scan or a magnetic resonance image (MR image). When the medical image data is to be displayed so that it can be viewed or recorded on a film as an image, each pixel value of the image data must be converted into a brightness using a predetermined non-linear conversion function. The conversion function is non-linear because the dynamic range of the medical image is wider than that of a display screen. If all the pixel values included in the pixel range of the digital image data are converted into a brightness using a linear conversion function, a clear image cannot be displayed. This conversion function is called the display window, which has the window level and the window width as parameters.

In a conventional system, a user manually adjusts the window level and width while viewing the image displayed on the screen, and sets the display window (conversion function) that enables display of a necessary portion in the image with optimum contrast and brightness. This operation becomes a burden for the user if the user adjusts the window level and width for a number of images at a time.

In a magnetic resonance imaging device or a computed tomography device, an attempt is therefore made to predetermine standard values for the window level and the window width for each scanning condition and to read out the corresponding preset standard values depending on the scanning condition at each scan in order to automatically adjust the window level and the window width.

It has also been proposed to set the display window depending on the characteristics of the image data. An example of such an apparatus is the image display apparatus disclosed in Japanese Patent Disclosure No. 63-84526. This apparatus forms a histogram as shown in Fig. 1A with respect to a pixel value from the image data and calculates a threshold value Th based on the maximum frequency Hmax, for example, 0.05 x Hmax, thereby setting a range (Wl to Wh) having a frequency of level Th or more. The apparatus sets the window width WW (Wl to Wh) and the window level WL (= (Wh + Wl)/2) for displaying the image based on the lower pixel value limit we and the upper pixel value limit Wh as shown in Fig. 1B, thereby generating the conversion function or display window for converting the pixel values into the brightness.

However, this conventional arrangement has the following problem. That is, since the display window is set based on only the threshold value obtained from the maximum frequency in the histogram, it often does not coincide with a display window set by a user. If the automatically set window is not suitable for observation, the user must adjust the automatically set window, resulting in a troublesome operation.

EP-0 312 596 A discloses a digital image display apparatus in which only an image quality is calculated based on histogram data. In particular, the window level of a display window is calculated based on the approximate average of the pixel data. The window width is calculated based on the obtained window level.

It is an object of the present invention to provide a digital image display apparatus which can convert a pixel value of digital image data into a brightness using an appropriate conversion function to thereby display a clear image.

A digital image display device in accordance with the present invention has the features defined in claim 1 or 11. Further features are defined in the subclaims.

According to the present invention, the image quality which indexes the clearness of the display image is defined, and the display window is set using this image quality. Therefore, a display window which results in a clear image can be easily set.

Additional objects and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The objects and advantages of the invention may be realized and obtained by means of the features and combinations particularly pointed out in the appended claims.

This invention can be better understood from the following detailed description in conjunction with the accompanying drawings:

Fig. 1A and iB show graphical representations illustrating a principle of conventional, automatic adjustment of a window,

Fig. 2 is a block diagram showing a first embodiment of a digital image display device according to the present invention,

Fig. 3 shows a flow chart in which a schematic operation of the first embodiment is illustrated,

Fig. 4 shows a histogram of pixel values generated in the first embodiment,

Fig. 5 shows a relationship between the pixel value and the brightness of the image displayed by a given display window,

Fig. 6 shows an evaluation function relating to a brightness value MVPG which corresponds to the picture element value MVP as a first estimated value to be used for calculating an image quality in the first embodiment,

Fig. 7 is a flowchart showing processing for calculating an optimal display image based on image quality in the first embodiment using a hill climbing method,

Fig. 8 is a diagram showing a change in the display window which is used to detect the change in the image quality depending on the change in the display window using the flow chart shown in Fig. 7,

Fig. 9 shows a graphical representation illustrating the principle of the hill climbing method,

Fig. 10 is a block diagram showing a second embodiment of a digital image display device according to the present invention,

Fig. 11 is a flow chart illustrating a schematic operation of the second embodiment,

Fig. 12 is a flowchart showing in detail the processing for calculating an optimal display window in the flowchart shown in Fig. 11,

Fig. 13 is a block diagram showing a configuration of a neural network in the second embodiment,

Fig. 14 is a flowchart showing in detail the processing for calculating the image quality in the flowchart shown in Fig. 12,

Fig. 15 is a view illustrating the learning of data in the neural network according to the second embodiment,

Fig. 16 is a block diagram showing a third embodiment of a digital image display device in accordance with the present invention,

Fig. 17 is a flowchart showing a schematic operation of the third embodiment,

Fig. 18 shows a histogram of pixel values generated in the third embodiment,

Fig. 19 shows a principle of peak value detection,

Fig. 20 shows the importance of the peak value,

Fig. 21 is a flowchart showing the processing for determining a display window in the third embodiment. Fig. 22 is a flowchart showing the display of image data in the third embodiment.

Fig. 23 is a flow chart showing a principle of a fourth embodiment of the digital image display apparatus according to the present invention,

Fig. 24 shows a contrast scale,

Fig. 25 shows a flow chart showing a schematic operation of the fourth embodiment,

Fig. 26 shows a neural network included in a digital image display device according to a fifth embodiment of the present invention,

Fig. 27 is a flow chart showing a schematic operation of the fifth embodiment,

Fig. 28 shows a table of the final state or the final condition γ T,

Fig. 29 is a flowchart showing processing for calculating the image quality in the fifth embodiment,

Fig. 30 shows the learning of data in a second-stage neural network or a two-stage neural network in the fifth embodiment,

Fig. 31 shows a relationship between an evaluation value and an evaluation result by the user, which is used to evaluate the display window set according to the present invention, and

Fig. 32 shows a relationship between an evaluation value and image quality

Embodiments of a digital imaging device according to the present invention will be described with reference to the accompanying drawings. Fig. 2 is a block diagram showing a configuration of a first embodiment of the present invention. Referring to Fig. 2, digital image data corresponding to, for example, a magnetic resonance image or a computed tomography image is stored in an image memory 1. The image stored in the memory 1 is displayed on a display 4. The display is connected to a window switch 3 for manually adjusting the window level and the window width and to a window controller 2 which serves to receive the image data from the memory 1 for automatically adjusting the window level, and the display window is both manually adjusted by the window switch 3 and automatically adjusted by the window controller 2 depending on the properties of the image data to be displayed.

An operation of the first embodiment will be described below. Fig. 3 is a flow chart showing a schematic operation of the window control device 2. In a step #110, a frame of the image data or a single image is read out from the image memory 1. This image data is supplied to a display 4. The digital image data contains a plurality of picture elements (pixels), each of which has a picture element value represented by the digital data. The pixel values of the image data are converted into brightness, and the image is displayed on the display 4 depending on the brightness. A function used to convert the pixel value into brightness is called the display window, which includes the display width and the display level as parameters. Therefore, the same image data is displayed with different brightness and contrast when the display window changes. At a step #120, a histogram is generated from the image data, the histogram indicating the frequency of pixel values included in the read-out image data. At a step #130, candidate values for the window level and the window width are set. At a step #140, an image quality is calculated which represents an index of the clarity of the image displayed on the display 4 and which preferably indicates the clarity of a specific portion of the image to be viewed when the image data is displayed with a given window level and a given window width.

At a step #150, either the window level WL or the window width WW, or both of these, are changed by ± ΔWW. The amount of change is not limited to this value. At a step #160, the image quality obtained when the image data is displayed based on the changed window level and width is calculated. At a step #170, the image quality obtained by the candidate display window calculated at the step #140 is compared with the image quality obtained based on the display window shifted from the candidate by ± ΔWW and calculated at the step #160. The candidate values for the window level and width are changed by 1 ΔWW, or the amount of change is decreased, depending on a difference between these two image qualities. This processing is repeatedly performed while ΔWW is changed. WW is reduced until the image quality obtained based on the candidate display window becomes maximum. The candidate display window that corresponds to the maximum image quality represents an optimal display window. This means that the optimal display window is determined by the hill-climbing method or the approximation method. In this way, the optimal display window can be automatically calculated depending on the overall characteristic of the histogram of each image. If the image displayed by the display window calculated by the window controller 2 is not clear, a user can finely adjust the window level and the window width using the window switch 3.

Each of the steps shown in Fig. 3 will be explained in detail below. Fig. 4 shows an example of the histogram generated in step #120, assuming that the digital image is constituted by a magnetic resonance image (MRI image). Here, the pitch of the histogram is set to one pixel value (MRI image value). However, the pitch may be freely set, for example, to four MRI image values. Since the histogram shown in Fig. 4 is obtained directly from the image data and includes a background portion, this background portion is preferably removed from the histogram. In general, in a MRI image of a human body, a part of the image corresponding to a portion outside the body has a small and substantially constant MRI image value and need not be considered. Therefore, first, an MRI value "A" corresponding to a boundary between the background portion H and a portion in the human body in the histogram is calculated. For this purpose, a peak frequency Hmax of the histogram within a predetermined range is calculated from a minimum MRI value. If the MRI value at the peak is assumed to be Hmax = h, the peak point is represented by (h, Hmax). A plurality of points, for example, five points (h + pH, H1), (h + 2 x pH, H2), (h + 3 x pH, H3), (h + 4 x pH, H4), and (h + 5 x pH, H5) are determined. Here, pH denotes a pitch of the histogram. A line ln connecting these points is calculated by the least squares method. A magnetic resonance image value at an intersection between the line Ln and an axis of the histogram that represents the frequency of zero is calculated and defined as the value "A", and a portion containing smaller magnetic resonance image values than the magnetic resonance image value "A" is set as the background H. These processings are executed at step #120.

The image quality calculated at step #140 will now be described. In this case, the image quality Q defined as follows is calculated:

Here, Vn denotes an estimated value of a feature n (n = 1 to k), and Wn denotes the weight coefficient of the estimated feature n. The weight coefficient Wn represents an empirically determined correlation ratio of the estimated feature n with respect to the image quality Q, such as W1 = 5, W2 = 4, W3 = 3, W4 = 2, and W5 = 2. The estimated value Vn is obtained using an estimation function Fn defined for each estimated feature. It is difficult in some cases to initially set the optimal values Wn and Fn. In this case, a user can set these values to be optimal by observing the display image using the display window.

Even when the estimation feature includes a brightness value, a contrast, a brightness balance, and the like, an estimation function fl is described below, which concerns a brightness value corresponding to a pixel value MVP. In this case, V1 = F1 (MVP).

As shown in Fig. 5, each magnetic resonance image value of the image data is converted into a brightness that changes continuously within the window width WW. Note that all magnetic resonance image values below WL - (WW/2) are displayed as the darkest, and all magnetic resonance image values above WL + (WW/2) are displayed as the brightest. Referring to Fig. 5, the brightness (256 gradation levels) is expressed as a value that varies continuously and linearly between the gray level values 0.5 to 16.5, which are shown on the ordinate for convenience. Assuming that the magnetic resonance image value is X and that the gray level value is Y, X and Y for a straight section within the window width WW are given as follows:

Y = 8.5 + 16 x (X - WL)/WW ...(2)

X = WL + WW x (Y - 8.5)/16 ... (3)

The pixel value MVP is defined as the magnetic resonance image value of the most desired portion within a frame of the image. Namely, the MVP is the magnetic resonance image value at a peak value P in a portion of the histogram containing the magnetic resonance image values greater than "A" (Fig. 4). This is because it is assumed that the entire image can be clearly seen if the image is displayed so that the magnetic resonance image value having the highest frequency is clearly visible. In this case, the brightness for a clear image is empirically judged to be slightly higher than the average brightness. Therefore, the evaluation function F1 is set in such a way that the estimated value V1 is maximized when the brightness level (gray scale) corresponding to the value MVP is equal to 10, and is decreased when the brightness corresponding to the value MVP has a larger difference from or than 10. The evaluation function is not defined linearly but is fixed nonlinearly because the human senses are ranked similarly to a membership function used in fuzzy theory. Fig. 6 shows an example of the evaluation function F1. In Fig. 6, the ordinate indicates a function value F1, and the abscissa indicates the brightness value MVPG corresponding to the value MVP.

Other evaluation functions are set in a similar manner. In this embodiment, the total of five evaluation features is set. The second feature is the brightness value, which corresponds to the value MRI "A". Since the magnetic resonance values, which are smaller than the value "A" corresponding to the background H, they are preferably dark. Therefore, an evaluation function F2 is defined such that a function value V2 is maximized when the brightness value corresponding to the magnetic resonance image value "A" is equal to 0.5, and is decreased when the brightness value corresponding to the magnetic resonance image value "A" is increased.

The third feature is an area of a portion of the image having a brightness value of 12. The portion of the brightness value 12 is assumed to be a portion having brightness values of 11.5 to 12.5 as shown in Fig. 5. Therefore, when Y = 11.5 to 12.5 is substituted into the equation (3), a range for the magnetic resonance image value X₁₂ corresponding to the brightness value 12 in the histogram can be represented as follows:

X₁₂₂ = (3 x WW/16) + WL to (WW/4) + WL ... (4)

The area corresponding to the brightness value 12 is therefore the area of the histogram containing the pixel values that fall within the range of equation (4).

The fourth feature is an area of a portion having the highest brightness level in an image. In Fig. 5, the highest brightness level is 16.5. Therefore, if Y = 16.5 is substituted into the equation (3), a magnetic resonance image value X16.5 corresponding to the brightness value 16.5 in the histogram is represented as follows:

X16.5 = (WW/2) + WL ... (5)

The area corresponding to the brightness level 16.5 is therefore the area of the histogram that contains pixel values that are larger than the value of equation (5).

The fifth feature is a brightness balance, which is defined as follows:

Brightness balance = (area containing brightness values from 4 to 8)/(area containing brightness values from 9 to 13) ... (6)

Each area of the numerator and denominator of equation (6) can be obtained in a similar way as the third and fourth features.

Next, the processing at step #170 for determining the optimum display window will be described. Fig. 7 is a flow chart showing this process in detail. At step #210, initial values WLo and WWo of the window level and the window width are calculated. These initial values are the first candidates for the window level and the window width. These initial values are set so that WLo = MPV is satisfied so that the value MVP is located at the center of the window width as shown in Fig. 4, and WWo = 2 x (MPV - A) is satisfied so that the value "A" coincides with the lower limit of the window width so that the entire image except the background H can be clearly seen. At step #220, the above-mentioned initial values are set as the candidate values WLc and WWc for the window level and the window width (WLc = WLo, and WWc = WWo). At a step #230, WWo/2 is set as an initial value for a change width ΔWW for the display window for a hill climbing method described later.

At a step #240, an image quality Qc of an image displayed when the window level and the window width are set to be equal to WLc and WWC, and the image qualities Qd1 and Qd8 of images displayed when the window level or the window width or these two sizes are changed by ΔWW as shown in Fig. 8 are calculated in accordance with the equation (1). In this case, the image quality Qd1 is obtained at a window level WLc + ΔWW and a window width WWc, the image quality Qd2 is obtained at a window level WLc + ΔWW and a window width WWc + ΔWW, the image quality Qd3 is obtained by the window level WLc and the window width WWC + ΔWW, the image quality Qd4 is obtained at a window level WLc - ΔWW and a window width WWc + ΔWW, the image quality Qd5 is obtained with a window level WLc - ΔWW and a window width WWc, the image quality Qd6 is obtained with the window level WLc - ΔWW and a window width WWc - ΔWW, the image quality Qd7 is obtained with a window level WLc and a window width WWc - ΔWW and the image quality Qd8 is obtained with the window level WLc + ΔWW and a window width WWC - ΔWW. It is not necessary to make the amount of change in the window level the same as that of the window width. As an example, it is possible to determine Qd1 to Qd8 at the following values respectively (WLc + Δ WW, WWc), (WLc + Δ WW/2, WWc + Δ WW), (WLc, WWc + 2 x Δ WW), (WLc - Δ WW/2, WWc + Δ WW), (WLc - Δ WW, WWc), (WLc - Δ WW/2, WWc - Δ WW), (WLc, WWc -2 x Δ WW), (WLc + Δ WW/2, WWc - Δ WW).

At step #250, the maximum image quality Qd is detected from the image qualities Qd1 to Qd8, and the window level WL and the window width WW corresponding to the detected image quality Qd are detected as WLd and WWd, respectively. At step #260, the image quality Qc of the candidate values of the display window is compared with the image quality Qd of the display window slightly shifted from the candidate values to check whether Qc is smaller than Qd. If the result at step #260 is yes, the values of the display window corresponding to the image quality Qd are more suitable than the candidate values for the display window. Therefore, at step #270, WLc = WLd, WWc = WWd, and Qc = Qd are set. This is essentially equivalent to a procedure in which the candidate values for the window level and the window width are changed to those corresponding to the image quality Qd, as shown in Fig. 9. The amount of change ΔWW is multiplied by β at a step #280, and the flow returns to the step #240. In this case, β is a constant satisfying the condition 1 < β, so that the amount of change is increased.

If the result in step #260 is no, it is determined that the candidate values are suitable. In order to determine the size of the change when repeating the above Therefore, in order to reduce the amount of change ΔWW specified in the judgment, the amount of change ΔWW is multiplied by α at a step #220. In this case, α is a constant satisfying the conditions 0 < α < 1 and 0 < α x β < 1. At a step #300, it is checked whether ΔWW ≥ γ. In this case, γ is a constant used to determine the end of the processing. If the result at step #300 is YES, the flow returns to step #240 to calculate the image quality based on the candidate values and the image qualities based on the eight windows around the candidate as shown in Fig. 8. If the result at step #300 is NO, the current candidate values WLc and WWc are set as the finally obtained appropriate window levels and widths WL and WW, and the processing is terminated.

According to the first embodiment described above, an index indicating the clarity of the display window that satisfies the human sense is defined as the image quality depending on the various evaluation items obtained from the histogram of the pixel values of the image data, and the image quality obtained based on the candidate for the display window is compared with the image quality obtained based on the window level and a window width that are slightly shifted from the candidate values. Depending on the comparison result, this comparison is performed while shifting the candidate values and the amount of shift is gradually reduced, thereby setting an optimum display window by the so-called hill climbing method using the image quality. Therefore, an optimum display window can be easily obtained for each image. As a result, a burden on the user can be reduced.

Although the evaluation function and the weight coefficient are manually set and determined in the above embodiment, these values can also be determined automatically. For example, the weight coefficient Wn in the following manner. In this case, if it is assumed that the number of features is 5 and that Wn is an integer from 1 to 5, 3125 (5⁵) combinations can be set for W1 to W5. A plurality of standard images, for example, 10 standard images are displayed, and the window level and window width set by an expert user (a radiologist when the image is a magnetic resonance image) are stored for each image. Display windows corresponding to the 3125 combinations of W1 to W5 are compared with the display window set by the expert user, and those values corresponding to the next display window are set as W1 to W5. Similarly, an optimum value for Fn can be set such that a difference with respect to the display window set by the expert user is minimized. The above processing can be realized not only by a computer with a normal procedure-oriented program, but also by an expert system using rules. Since the clarity of the image depends largely on the human senses, an expert system is more preferable.

Next, a second embodiment of the present invention will be described. Fig. 10 is a block diagram showing the second embodiment. The second embodiment is the same as the first embodiment shown in Fig. 2 except that a neural network is connected to a window controller 2. In the first embodiment, the image quality is defined as a function and is obtained by a calculation process.

In the second embodiment, however, the image quality is determined by the neural network 5.

Fig. 11 shows a schematic operation of the second embodiment. In a step #320, a single frame of the image data is read out from the image memory 1. In a step #330, a histogram is generated from the image data, which has a frequency or frequency of picture element values. This histogram is similar to the histogram in the first embodiment shown in Fig. 4. At a step #340, the value "A" is obtained as in the first embodiment. At a step #350, the value MVP is obtained as in the first embodiment. At a step #360, the window level and the window width appropriate for the image (for example, the value "A" and the value MVP) are set. This processing will be described in more detail later. At a step #370, the image data is displayed by the set window level and window width. At a step #380, a user observes the displayed image and, if necessary, adjusts the window level and width using the window switch 3 to achieve fine adjustment.

Each step shown in Fig. 11 will be explained in detail below. Fig. 12 shows the determination of the appropriate display window at step #360 in more detail. In a similar manner to the first embodiment, the hill climbing method is used as the method for determining the display window. However, different from the first embodiment, the image quality with respect to the display window is not calculated by a calculation process but is obtained by the neural network 5. That is, at step #410, candidate values for the window level and the window width are determined. At step #420, the image quality of the image data displayed with the specified window level and the specified window width is determined by the neural network 5. At step #430, the window level or the window width or both are changed by ΔWW. At step #440, the image quality for the image data displayed with the changed window level and the changed window width is determined. At a step #450, the image qualities obtained at steps #420 and #440 are compared with each other, and candidate values for the window level and the window width are changed depending on a difference between the two image qualities. In addition, optimal values for the window level and the width are determined while gradually decreasing ΔWW, the is when determined according to the hill climbing method. The processing for determining the optimum display window using the hill climbing method is the same as the processing in the first embodiment shown in Fig. 7.

As described above, the image quality in the second embodiment is determined using the neural network, whereas the image quality in the first embodiment is calculated using a function such as a membership function used in a fuzzy theory. Therefore, the image quality suitable for the characteristics of each device or hospital can be easily obtained. As a result, the display window optimal for each device can be set by causing a neural network to learn depending on the device. Referring to Fig. 10, the window controller 2 transmits evaluation features (similar to those in the first embodiment) for calculating the image quality to the neural network 5, thereby obtaining the image quality. The neural network 5 generates the image quality based on the evaluation features and transmits the generated image quality to the window controller 2. As a result, the window controller 2 determines the image quality.

The neural network 5 is described below. Fig. 13 shows a configuration of the neural network 5. The network 5 is formed by three layers consisting of an input layer 7, an intermediate layer 8 and an output layer 9. The intermediate layer 8 may consist of a plurality of layers. The input layer 7 contains five units 7a to 7e. The five evaluation feature data explained in the first embodiment are fed into the neural network 5. This means that the brightness value MVPG, the brightness value AG, the area of the brightness value 12, the area of the maximum brightness and the brightness balance are fed into the units 7a to 7e, respectively. The number of input features is not limited to five. If the number of input features is increased or decreased, the number of units in the input layer 7 can be increased. be increased or decreased accordingly. A method for determining the input data is described in more detail below. In this embodiment, the intermediate layer 8 consists of 50 units and forms the network together with the units 7a to 7e in the input layer 7 and the output layer 9. The output layer 9 is formed by a unit and outputs the image quality with a value from 0 to 1. An input/output relationship of the unit contained in the neural network 5 can be represented by the following sigmoid function f(x):

f(x) = 1/{1 + exp(-x + b)} ... (7)

Here, b denotes the bias, which is a weight of a bias element (which has an output of "1") and which is determined by learning. x denotes the input variable. The neural network 5 must be made to learn in advance. This learning is described in more detail in the text below. In this case, it is assumed that the neural network 5 has finished learning.

Fig. 14 is a flow chart showing a sequence for determining the image quality using the neural network 5 in steps #420 and #440. Referring to Fig. 14, the brightness value MVPG corresponding to the value MVP of the image displayed with a given display window is obtained at step #510. The brightness value MVP can be represented as follows when X is replaced by MVP and Y by MVPG in the equation (3):

MVPG = 8.5 + 16 x (MVP - WL)/WW ...(8)

Since an input value for the neural network 5 must have a value from 0 to 1, the brightness value MVPG, which is calculated using equation (8), multiplied by a normalized coefficient of 1/16.5 to obtain a normalized value MVPG.

At step #520, the brightness value AG corresponding to the magnetic resonance image value "A" of the image displayed according to a given window is obtained. The brightness value AG is represented in accordance with the equation (3) in a similar manner to the brightness value MVPG as follows:

AG = 8.5 + 16 x (A - WL)/WW (9)

Also in this case, AG is calculated using equation (9) and multiplied by the normalized coefficient 1/16.5 to obtain a normalized value of AG.

At step #530, the area of a portion of the image having a brightness value at gray level 12 and displayed by the given window is calculated. More specifically, the area of a portion of the histogram at gray level 12 is calculated in accordance with equation (4), and the result is multiplied by a normalized coefficient of 1/100. However, if the normalized value exceeds 1, the value is regarded as 1.

At step #540, the area of the section having the highest brightness value, that is, the section at gray level 16.5 of the image displayed through the given window, is calculated. More specifically, the area of a section of the histogram is calculated in accordance with equation (5) and multiplied by a normalized coefficient of 1/30. If the normalized value exceeds 1, the value is considered to be 1.

At step #550, the brightness balance, that is (areas having brightness values from 4 to 8)/(area having brightness values from 9 to 13), is calculated as shown in of equation (6) for the image displayed through the given window. These areas are determined in a similar manner to steps #530 and #540. In this case, a normalized coefficient is equal to 1/6. If the normalized value exceeds 1, the value is considered to be 1.

At a step #560, the five values obtained at steps #510 to #550 are transmitted to the neural network 5 and input to the units 7a to 7e in the input layer 7. At a step #570, the image quality is obtained from the output layer 9 of the neural network 5.

The learning of the neural network 5 will be described below. As stated above, the neural network 5 must learn in advance in order to output an image quality suitable for the input evaluation feature data. Although a learning method is not limited to that in this embodiment, a back propagation method is used in this embodiment. That is, a predetermined number of learning data are repeatedly input to the neural network 5 to thereby perform learning by the back propagation method, and the learning is repeated until the error is reduced below a predetermined value. In this embodiment, the learning is carried out using the learning data given below. However, the method for generating the learning data is not limited to that in the present embodiment.

The learning data is formed based on a plurality of current image data and window levels and window widths set by an experienced user for displaying this image data. An image is selected and a window level WLG and a window width WWG are set for the image by the experienced user. A total of 25 combinations (WLS, WWS) are used as display windows for learning using a window level WLS of WLG - WWG/2, WLG - WWG/4, WLG, WLG + WWG/4, and WLG + WWG/2 and a window width WWS of WWG - WWG/2, WWG - WWG/4, WWG, WWG + WWG/4, and WWG + WWG/2. Five evaluation feature data for the brightness MVPG, the brightness AG, the area of brightness at gray level 12, the area of brightness at gray level 16.5, and the brightness balance are calculated for each of the 25 combinations of (WLS, WWS) at steps #510 to #550, and are input as input data for learning. Teacher data for these combinations of (WLS, WWS) are defined as shown in Fig. 15 to thereby form 25 sets of learning data. Subsequently, 25 sets of learning data are formed for each of an arbitrary number of images, and learning is repeatedly carried out using all the learning data until the values of the errors are reduced below a predetermined value.

According to the second embodiment described above, since the neural network 5 can perform a learning operation, a display window suitable for the characteristics of a single device or hospital can be automatically set without performing cumbersome operations such as setting an evaluation function and a weight coefficient as in the first embodiment. Furthermore, since the image can be displayed through the appropriate display window, the process of setting the display window to be performed for displaying the image is simplified, so that the burden on the user is reduced.

A third embodiment of the present invention will be described below. In the first and second embodiments, an optimum display window is determined for each image. However, an optimum display window varies even in an image depending on an object to be observed (a portion or the type of tumor), a purpose, or the like. In the third embodiment, therefore, an appropriate display window is determined for each object to be observed in an image, and the image is displayed using different display windows for different objects. Fig. 16 is a block diagram showing the third embodiment. A configuration of the third embodiment is the same as that of the first embodiment. that in the second embodiment, except that a selector switch 10 for selecting the object to be observed is connected to a window control device 2.

Fig. 17 shows a schematic operation in the third embodiment. Referring to Fig. 17, at step #5100, an image data is read out from the memory 1. At step #5300, a histogram indicating a frequency of pixel values is formed on the basis of the image data. This histogram has a plurality of peaks P1, P2 and P3 as shown in Fig. 18. At step #5300, the value "A" is obtained as in the first embodiment. At step #5400, the value MVP is formed for each object. As explained above, the value MVP is an MRI value of a portion which is considered to be most clearly desired by the user. However, since the number of portions to be observed is not limited to one portion in one image, the value MVP is calculated for each object. Suppose that MVPn are obtained by arranging MVPn in order of importance. MVPn are the magnetic resonance image values of a portion which is considered to be most desired by the observer with respect to an nth object. The object of observation differs between a case where an entire image is to be observed and a case where a specific object (for example, a portion or a tumor) is to be observed. In order to observe the entire image, it is assumed that the entire image can be clearly seen when the image is displayed such that a magnetic resonance image value having the highest frequency in the image excluding the background can be clearly seen. In order to observe the specific object (for example, a portion or a tumor), it is assumed that a peak of a partial histogram corresponds to a portion which is most desired to be observed. Therefore, in the case shown in Fig. 18, MVP1, MVP2 and MVP3 represent magnetic resonance image values of sections that are assumed to be most intense with respect to the respective objects. are to be observed. In this case, the number of observation objects m is 3. More specifically, the magnetic resonance image values MVPn are calculated by searching the peaks of the histogram located in the area above the magnetic resonance image value "A". As shown in Fig. 19, the peak is defined as a point having a frequency higher than that of several points on both sides of this point. As shown in Fig. 20, the importance of the peak is evaluated by considering, for example, (1) a distance xi between the peaks, (2) the height of the peak, and (3) the width W1, W2 of the peak. In this embodiment, the values MVP1, MVP2, and MVP3 are important in the order given. The observation objects (the number of MVP) are stored at the same time.

At step #5500, an appropriate window level and window width are set for each observation object (MVPn). This setting process is described in more detail below. At step #5600, the image data for each object is displayed using the window level and window width set for each object at step #5600.

At step #5700, a user observes the displayed image and makes fine adjustments of the window level and width using the window switch 3 if necessary.

According to the third embodiment, as explained above, the user's workload can be reduced because a display window suitable for the observation object can be easily set when the image is displayed.

Each step is described in detail below. The sequence for determining the display window suitable for each object, which is performed at step #5500, is explained with reference to Fig. 21. At step #610, a parameter n for designating the object is set to 1. At step #620, the display window suitable for an n-th object is determined. This display window can be determined from the same method as that in the second embodiment shown in Fig. 14. In this case, the evaluation features and the weight coefficients for the network are selected depending on the object (the type of MVP). The image quality need not be obtained by the neural network but may be calculated from a function as in the first embodiment. In this case, the evaluation features, the weight coefficient and the evaluation function for calculating the image quality are selected depending on the object (the type of MVP). As an example, W1 to WS are set to 5, 4, 3, 2 and 2 with respect to an object 1, and are set to 6, 4, 3, 2 and 1 with respect to an object 2. Note that the second evaluation feature for the object 1 is the brightness of value "A", but for the object 2 is the brightness of value "B", as shown in Fig. 2. The value "B" indicates an image element value at an intersection between the X axis and a line connecting a point located between a certain peak and an adjacent peak located to the left of the certain peak and having a lowest histogram value, and the certain peak. In this way, the evaluation function is selected depending on the object. The obtained window level and the obtained window width are stored in a window memory included in the controller 2. At a step #630, the parameter n is incremented. At a step #640, the parameter n is compared with the number of objects m. If n ( m , that is, if there remain objects whose display windows have yet to be determined, steps #620 and #630 are repeated. If n ) m , that is, if the display windows for all objects have been determined, the processing is terminated.

Fig. 22 shows the sequence of selecting a display window suitable for the object and displaying the illustrated step at step #5600. At step #650, the parameter n is set to 1. At step #660, the window level and window width corresponding to the n-th object are read from the window memory in the controller 9 and transferred to the display 4. As a result, the image is displayed so that the n-th object is most clearly visible. If the Selector switch 10 is pressed at step #670, the parameter n is incremented at step #680. At step #690, the incremented value of n is compared with the number m of objects already stored in advance; if n is equal to or smaller than m, the flow returns to step #660, and the image is displayed through the n-th display window. If n reaches the value of m, the flow returns to step #650, and n is reset to the value 1. This means that the display window for the first object is set again when selector switch 10 is pressed after the display window for the m-th object has been set. The display windows for the respective objects are set sequentially each time switch 10 is pressed during display of the image. Therefore, the user can display the image with the display window appropriate for the object by a simple operation.

According to the third embodiment, according to the above explanation, the optimal display window can be easily set for each observation object in an image.

A fourth embodiment of the present invention will be described below. Fig. 23 is a schematic view showing a sequence of image formation in the magnetic resonance imaging apparatus. At step #700, raw magnetic resonance image data is obtained. The obtained image data is corrected at step #710, and the corrected image data is reconstructed into an image using, for example, a two-dimensional Fourier transform at step #720. At step #730, a pixel value of the reconstructed image data is converted into a magnetic resonance image value using a contrast scale, that is, the reconstructed image is converted into a magnetic resonance image. The reconstructed image is a digital image formed by the pixels each having the pixel value and having the same matrix size as that of the magnetic resonance image. The pixel value of the reconstructed image changes within a wide range depending on the object and the scanning condition. The Magnetic resonance image is thus formed by the magnetic resonance image values within the range of -1000 to 2000 (or 4000) obtained by converting the pixel values of the reconstructed image using the contrast scale.

The contrast scale is generated as shown below:

Y = pX + q ... (10)

Here, Y denotes the magnetic resonance image value, X the pixel value of the reconstructed image, and p and q are the constants, respectively. In this embodiment, the constants p and q of the contrast scale are set depending on the reconstructed image so that the magnetic resonance image value is standardized. A block diagram of the fourth embodiment is the same as that of the third embodiment, and is therefore not shown.

A schematic operation in the fourth embodiment will be described below with reference to Fig. 25. At step #750, a reconstructed image data is read out from the image memory 1. At step #760, a histogram indicating a frequency of the pixel values is formed based on the reconstructed image data. As shown in Fig. 18, this histogram has a plurality of peaks P1, P2 and P3. The value "A" is formed at step #770, and the value MVP for each object is formed at step #780. At step #790, an appropriate window level and window width are determined for each object (each value MVPn). At step #800, the reconstructed image data for each object is displayed based on the set window level and window width appropriate for the object. At step #810, a user observes the displayed image and finely adjusts the window level and width using the window switch 3 if necessary. The processing described so far is exactly the same as that from steps #5100 to #5700 in the third embodiment shown in Fig. 17.

At step #820, as shown in Fig. 24, the constants p and q of the contrast scale (conversion function for converting picture element values into MRI values) are set based on the window level WL and the window width WW so that the upper limit WL + (WW/2) of the width of the display window is converted into MRI value U, and the lower limit WL - (WW/2) of the window width is converted into MRI value L. By substituting U WL + (WW/2) and L = WL - (WW/2) into the equation (10), p and q are calculated in the following manner:

p = (U - L)/WW ...(11)

q = {(U + L)/2} - {WL(U - L)/WW} ... (12)

In a step #830, the pixel value of the reconstructed image is converted into the MRI value using the calculated constants p and q of the contrast scale (conversion function) to thereby obtain the magnetic resonance image. In this case, most of the magnetic resonance image values of an effective portion (the portion to be observed) of the magnetic resonance image are included between the upper limit WL + (WW/2) of the display window and the lower limit WL - (WW/2). Therefore, by converting the pixel value of the reconstructed image into the magnetic resonance image values using the contrast scale, a magnetic resonance image is obtained in which the dynamic range of the magnetic resonance image value for a portion to be observed is between U and L, and the magnetic resonance image value is standardized to the range of the values to be observed. Since the magnetic resonance image value has no absolute meaning, it can be standardized in this way. In order not to limit the dynamic range of the magnetic resonance image, the range between U and L is preferably set as wide as possible.

At step #840, the magnetic resonance image is stored in the image memory 1. In order to display the stored magnetic resonance image at a brightness suitable for observation, the window level WL is set to (U + L)/2, and the window width WW is set to U - L. As a result, the same display window can be set for all images, and it is not necessary to set the display images for all respective images respectively. This means that it is sufficient to set the display window once because the magnetic resonance image values are standardized. The window setting process to be performed for displaying the image is simplified, thereby reducing the user's workload.

Furthermore, according to the present embodiment, it is also easy to adjust the display window to thereby set the constants of the contrast scale.

Although the display windows are set for the respective objects in the fourth embodiment, a display window for an image can be set without considering the objects in the image, as in the second embodiment. Furthermore, in the fourth embodiment, the image quality can be calculated using the evaluation function, as in the first embodiment.

A fifth embodiment of the present invention will be explained below. This embodiment relates to an improvement in the neural network. In the above embodiments, a neural network is used to determine the image quality. However, sufficient precision cannot be achieved if only one neural network is used. In the fifth embodiment, therefore, a number of neural networks are used sequentially to achieve sufficient precision.

The block diagram for the fifth embodiment is the same as that of the second embodiment shown in Fig. 10 and is therefore not shown. As shown in Fig. 26, the neural network 5 is formed by a network of two stages comprising a first neural network 5a and a second neural network 5b, wherein each respective evaluation feature data is input to the networks 5a and 5b. The two networks have different levels of accuracy with respect to the calculated image quality.

An operation of the fifth embodiment will be described below. A schematic flow is the same as that of the second embodiment shown in Fig. ii. Fig. 27 is a flow chart showing the processing in determining an optimal display window using the neural network. At a step #1010, initial values WLo and WWo are calculated for the window level and the window width. In the same manner as in the above-mentioned embodiments, the initial values WLo = MVP and WWo = 2x (MVP - A). At a step #1020, the initial values are set as the candidates for the display window such that WLc = WLo and WWc = WWo. At a step #1030, WWo/2 is set as an initial value for the amount of change ΔWW of the display window. At step #1040, 1 is set in a parameter NN indicating the stage of the network. At step #1050, an end condition value γ T(NN) for an NN-th stage of the neural network is taken out of a table and set to be equal to γ. Fig. 28 shows the table. In this embodiment, the search for the optimal display window is carried out by sequentially decreasing the amount of change. When the amount of change is decreased to be equal to a predetermined value or less, the search for the optimal display window is terminated. Therefore, the end condition value for the first stage of the neural network is the largest, and the other end condition values are smaller in the descending order.

At step #1060, the image quality is calculated using the NN-th stage of the neural network, and the optimal window level and window width for displaying the window are calculated depending on the image quality. In this case, the initial values for the candidate for the Display windows used in the processing performed by the first stage of the neural network are set at a step #1020 shown in Fig. 27, and the initial values for the candidate display window used in the processing performed by the subsequent stage of the neural network are the display window determined in the previous stage of the neural network. The display window is determined with a coarsest precision in the first stage 5a of the neural network, and it is set with finer precision in the second, third, and subsequent stages of the neural networks. This processing will be explained in more detail later.

At step #1070, the parameter NN is incremented to designate the next-stage neural network (next-stage neural network). At step #1080, it is determined whether the processing for all the neural networks has been executed. If the processing for all the neural networks has not been executed at step #1080, the flow returns to step #1050. If the processing for all the neural networks has been executed at step #1080, the display window obtained at that time is set as the optimal display window.

The processing for obtaining the image quality using the neural network 5 at step #1060 will be described in more detail below with reference to Fig. 29. At step #1110, the image quality Qc of the image displayed through the display window containing the candidate values WLc and WWc is calculated. This image quality Qc can be obtained in the same manner as in the second embodiment shown in Fig. 14. At step #1120, in a similar manner to the second embodiment, the window level or the window width or both of these sizes are changed by ± Δ WW with respect to the candidate values, and the image qualities Qd1 to Qd8 of the images displayed through the changed display windows are calculated. The amount of change is is not limited to the value ΔWW. At a step #1130, a maximum image quality Qd is determined from the image qualities Qd1 to Qd8, and the window level WL and the window width WW corresponding to the image quality Qd are set as WLd and WWd, respectively. At a step #1140, the image quality Qc obtained with the candidate values for the display window is compared with the image quality Qd obtained with display window values slightly shifted from the candidate values to check whether Qc is smaller than Qd. If the result at step #1140 is yes, the display window corresponding to the image quality Qd is more suitable than the candidate for the display window. Therefore, at a step #1150, WLc = WLd, WWc = WWd, and Qc = Qd are set. That is, this corresponds to an event where the candidate values for the window level and the window width move to the values corresponding to the image quality Qd as shown in Fig. 9. At step #1160, the amount of change ΔWW is multiplied by β, and the flow returns to step #1120. In this case, 1 ≤ β, that is, the amount of change is increased.

If NO in step #1140, the candidate values are set as appropriate values, and it is checked at step #1170 whether ΔWW is larger than y. In this case, γ is a constant used to determine the end, which is read from the table shown in Fig. 28. If NO in step #1170, it is determined that the candidate values WLc and WWc are the appropriate window level WL and the appropriate window width WW and are finally set, and the processing is terminated. If YES in step #1170, the amount of change ΔWW is multiplied by α in step #1180 to thereby perform a determination with a smaller amount of change. In this case, α is a constant satisfying the conditions: 0 < α < 1 and 0 < α x β. < 1. At step #1190, it is checked whether Δ WW is smaller than γ or not. If the result is no at step #1190, the flow returns directly to step #1120. If the result is yes at step #1190, γ is set as the amount of change Δ WW at step #1200, and the flow returns to step #1120. In this way, in the fifth embodiment, an optimum display window can be obtained based on the image quality by using a hill climbing method.

The learning by the neural networks 5a and 5b will be described below. The neural network 5a of the first stage performs a learning operation as already described in the second embodiment. However, in the neural networks of the second and subsequent stages, in order to further improve the accuracy, a sampling pitch for the display window in the learning is set more finely in a section having high image quality so that the networks learn different learning data. For this reason, the neural network of the second stage learns by a fine pitch only in a section close to WLG and WWG set by an experienced user. It is to be noted that if a single network were made to learn using a fine pitch, not only would the learning efficiency be reduced, but also errors in learning could not be reduced because a dynamic range of the image is limited in a portion close to the set values WLG and WWG. Therefore, this learning method is not practical. The learning method in the second and subsequent stage networks is the same as that in the first stage network, but training data is formed, which will be explained later. However, the method for generating the training data is not limited to that in the present embodiment.

An image learned by the first-stage network is selected, and a total of 25 points (WLS, WWS) are sampled as the display windows for the selected image using WLG -WWG/4, WLG - WWG/B, WLG, WLG + WWG/3, and WLG + WWG/4 as the window level WLS, and WWG - WWG/4, WWG - WWG/8, WWG, WWG + WWGIB, and WWG + WWG/4 as the window width WWS. In a similar manner to the second embodiment, five evaluation feature data, that is, the brightness value MVPG, the brightness value AG, the area having the brightness value 12, the area having the brightness value 16.5, and the brightness balance for each of the 25 windows (WLS, WWS) is calculated at steps #510 to #550 shown in Fig. 14 and used as the input data for learning. Teacher data with respect to these windows (WLS, WWS) are defined as shown in Fig. 30 so that 25 sets of learning data are formed. Subsequently, the 25 sets of learning data are formed for an arbitrary number of images. All the formed learning data are repeatedly learned until the value of the errors is reduced below a predetermined value.

When a plurality of neural networks are used, the image qualities produced by the respective networks are generally different even for the same display window, so that a problem of so-called mismatch arises. However, in this embodiment, no such problem occurs because the neural networks of the plurality of stages are sequentially used for the image quality production and the search for the optimum display window is independently performed by using an immediately preceding search result as an initial value for the next search. Although the number of stages of the networks is 2 in the illustrated embodiment, a system having three or more networks can also be used in practice in a similar manner by setting a finer sampling pitch.

According to the fifth embodiment, as described above, a plurality of neural networks learn data with a coarse precision and data with a fine precision, and the networks are used step by step to thereby obtain an optimal display window. Therefore, the display window can be adjusted for a clearer image. In the fifth embodiment, a plurality of networks having the configuration according to the second embodiment are used. However, a plurality of networks having the configuration according to the third or fourth embodiment may also be used.

In the following, teacher data for use in the neural network shown in Fig. 15 or Fig. 30 will be described. The teacher data represents an index indicating the clarity of the image displayed by means of a given window. It is necessary to define the teacher data to indicate the clarity. Therefore, the difference between the brightness of an image displayed by a reference display window, where the user feels that the image is displayed most clearly, and a brightness of an image displayed by the given display window is evaluated, and the teacher data is formed based on the evaluation. The image quality Qe of the image displayed by the given window is therefore defined as follows:

Qe = - (K1 x EW) + K2 ... (13)

Here, K1 and K2 denote the normalized coefficient and the normalization coefficient, respectively, and EW denotes the evaluation value for evaluating a difference between the image displayed by the reference display window and that displayed by the given display window.

EW is defined according to the equation given below. This equation represents a weighted average per pixel of a square sum of the difference between the brightness (gray scale) of the image displayed by the reference display window and the brightness (gray scale) of the image displayed by the given display window. Assuming that W(GSr) = 1 in the equation below, the numerator of the equation represents a square sum of a difference between the gray scales of the pixel values. Since the numerator represents the number of pixels, the size (numerator) 1 (denominator) indicates the average per pixel. However, the average brightness in an image is important, and a shift in the maximum and minimum brightness values exerts a considerable influence. The weighting is therefore carried out with respect to the reference brightness:

(EW)² = {&Sigma; W(GSr(x, y)) x (GSs(x, y) - GSr(x, y))²} all image elements

/ &Sigma; W (GSr(x, y)) all image elements ...(14)

Here, GSR(x, y) denotes the grayscale value of a pixel (x, y) with respect to the reference window, while GSs(x, y) denotes the grayscale value of the pixel (x, y) with respect to the given window and W(GSr(x, y)) denotes the weighting coefficient in the grayscale or the grayscale value GSr.

When the gray scale is represented by values from 0.5 to 16.5, a gray scale GS obtained when a picture element value PV(x, y) is displayed through a window (WL, WW) is given as follows:

GS = 0.5 (if PV &le; WL - (WW/2))

GS = 16 x (PV - WL)/WW + 8.5 (if WL - (WW/2) < PV < WL + (WW/2))

GS = 16.5 (if WL + (WW/2) ≤ PV) ... (15)

W(GSr) is defined as follows:

W(GSR) = 0.1 (if GSr &le; 0.5)

W(GSR) = (GSr - 0.5) x 0.9/6.5 + 1 (if 0.5 < GSR &le; 7.0)

W(GSR) = 1.0 (if 7.0 < GSr ≤ 10.0)

W(GSR) = 0.1 - (GSr - 16.5) x 0.9/6.5 (if 10.0 < GSr ≤ 16.5)

W(GSR) = 0.1 (if 16.5 < GSr) ... (16)

Equation (14) can be rewritten as follows if the histogram of pixel values is used:

(EW)² = {&Sigma; W(GSr(PV)) x (GSs(PV) - GSr(PV))² x H(PV)} all PVs

/ &Sigma; {W(GSr(PV) x H(PV)} all PVs

Here, PV denotes the pixel value and H(PV) denotes the frequency of the pixel value PV.

The image quality Qe is defined in the manner described above. It is necessary to show that the image quality Qe corresponds to the human perception. Since Qe is a value normalized by EW, EW is compared with the human perception. This comparison method takes place in the following way:

(1) 12 magnetic resonance images of three patients were prepared.

(2) A radiologist sets an optimal viewing window and the reference window is determined for each image.

(3) Six to seven display windows are set for each image. These display windows are set so that EW is varied as widely as possible.

(4) EW is calculated in each set window.

(5) An image is displayed in each set window and evaluated by the radiologist. The evaluation results are shown in Fig. 31. Referring to Fig. 31, the evaluation is represented as follows:

(1) A display window is suitable and the picture is clear: ...

(2) A display window is tolerable, but fine-tuning is required: ...Δ

(3) An object to be scanned can be recognized, but is unclear: ...X.

As can be seen from Fig. 31, EW is consistent with human perception.

If only WL of the adjusted window is shifted by one grayscale compared to the reference window, EW = 1 is obtained, since all values of GSs(x, y) - GSr(x, y) become 1. It is claimed that a person is able to recognize a difference of one gray level (grayscale or grayscale value). In other words, a person is not greatly affected by a difference within one gray level. This also explains the experimental findings.

Therefore, if the neural network learns the image quality Qe defined in equation (13) as the teacher data, it is possible to learn the relationship between the input to the neural network and the clarity of the image.

In this case, the normalization coefficients K1 and K2 are set in the following manner. Namely, since the teacher data is 0 to 1, K2 is set to 1. Since the system compares the image qualities to find the highest image quality, the image quality may have a relative value. Since the range of EW is wide when the range for the learning window is wide (that is, when the sampling is coarse), K1 is set to be approximately 1/10 for the first-stage neural network. When the range for the learning window is small, K1 is set to be approximately 1/3 for the second-stage network to thereby broaden the dynamic range. In any case, Qe is set to 0 when a negative value is obtained in equation (13).

The relationship between the image quality Qe according to equation (13) and that according to equation (1) will now be described. The image quality Qe defines the clarity of the image based on the brightness of the image. As described above, it can be shown that the image quality Qe conforms to human perception. It has therefore been assumed that Qe is correct. The equation (1) determines the clarity of the image based on the histogram of the pixel values. If the correlation between Q and Qe is large, it is possible to obtain the optimum display window in the first embodiment based on the image quality Q. In other words, it is necessary to tune the system so that the correlation between Q and Qe is enhanced. Since the image quality has a relative value, the normalized coefficient or standardization coefficient of equation (13) does not need to be determined, but the correlation between Qe and Q is required. For this purpose, a minus correlation or negative correlation must exist only between Q and EW. In other words, Q must be adjusted or the tuning must be carried out in such a way that the correlation between Q and EW is improved.

Q was calculated based on the currently set values Fn and Wn and compared with EW. Fig. 32 shows the comparison results for one image. As can be seen from Fig. 32, this correlation coefficient is equal to -0.89, that is, a negative high correlation is obtained.

Although the object of the present invention is to set the window close to the reference window, EW can be used to evaluate the display window set by the existing system. In this case, EW is calculated with respect to the display window set by the system. If the calculated EW is small, it can be concluded that a display window is set that is close to the reference window. If the calculated EW is larger, it can be concluded that a display window is set that is far from the display window. In this case, the value of EW in the be considered in the following manner, as explained above with regard to the experimental results.

If EW ≤ 1, the display window is suitable and the image can be seen clearly.

If 1 < EW ≤ 2.5, the display window is tolerable, but fine-tuning is required.

If 2.5 < EW, the object can be recognized but can only be viewed unclearly.

The present invention is not limited to the above embodiments, but may be modified in various ways. Although the image described in the above embodiments is a magnetic resonance image, the image may be a digital image only, but is not limited to a medical image. Even if a medical image is used as the image, the image may be a CT image and may be obtained by analog-to-digital conversion of an analog image such as a nuclear medicine image or a fluoroscopic X-ray image.

Modifications of the present invention are described below.

In the above embodiments, both the display window level and width are changed to detect a change in image quality. However, one of the display window levels and widths may be set as the candidate value while the other is changed.

In the above embodiments, the histogram is formed for each image and optimum values for the display window are set for each image. However, all images belonging to the same group can be displayed using the the same display window. For example, the histogram is formed from a plurality of slice images obtained during a volume scanning operation, and the display window is set based on the histogram, thereby viewing a plurality of images using the same display window. In this case, the images of the same group are viewed using the same display window. In addition, the histogram from a plurality of images can be used to form the teacher data for the neural network. Images obtained continuously under the same scanning conditions are preferably viewed using the same display window. Therefore, a display window selected by a user for a first image (if fine adjustment has been performed, the display image obtained after the fine adjustment) and the corresponding image quality are stored. If display of a second image is designated, the image quality of the second image is calculated with respect to the display window for the first image and compared with the image quality of the first image already stored in advance. If a difference between the two image qualities is equal to or smaller than a predetermined value, the display window is not changed so that the second image is displayed with the display window for the first image. If the difference is larger than the predetermined value, the display window for the second image is determined in accordance with the present invention and the image is displayed with the determined display window. In order to view images obtained continuously under the same sampling conditions with the same display window, the display window may be determined with respect to the typical image (for example, an image at a central position) of the same group so that all images of the same group are displayed with the same display window.

Although the display window in the embodiments explained above is linear within the width of the display window, it may also be non-linear.

In the third embodiment, the magnetic resonance image value that is assumed to be an object of observation is searched, and the display window that has the highest image quality in terms of the magnetic resonance image value is determined. However, in order to simplify the calculations, the image qualities may be calculated based on the subsequent display windows. The display windows are arranged from the one corresponding to the highest image quality, and the image is displayed through the display window from the one corresponding to the highest image quality and the one corresponding to the image quality in the decreasing order each time the selector switch 10 is pressed. In this case, if a user finds a comparatively suitable display window, he or she must operate the window switch 3 to finely adjust the display window manually. The display windows are selected in the following manner:

WL = LL - DD/4, LL - DD/8, LL, LL + DD/8, LL + DD/4 WW = 3DD/4, 7DD/8, DD, 9DD/8, 5DD/4

Here, LL and DD are constant.

In the third embodiment, an observation object is estimated based on the peak value of the histogram. However, the pixel value of the portion to be observed may be estimated using the following information or method.

(1) A method for performing simple image recognition to find a portion where pixel values are relatively close to each other.

(2) Sampling conditions.

(3) Slices or sections (an axial section, a coronal section, and a sagittal section).

(4) Scanning object that can be obtained through a hospital filing system.

Features for calculating the image quality, the evaluation function, the weight of the evaluation function, the weight of each unit of the neural network can be selected depending on the above-specified information and the above-specified features (for example, a total area, a peak position, a shape and a height) of the histogram.

With the third embodiment, the display windows for all objects are already calculated in advance, and the image with the calculated window suitable for the object is displayed when the selector switch 10 is pressed. However, the display window suitable for the object may be calculated each time the selector switch is pressed.

Modifications for determining the image quality using the neural network are described below. Although the same weight coefficients are used for all images in the above-described embodiments, the weight coefficient may be selected depending on the type of the image. In this case, the weight coefficient is transmitted from the window controller to the neural network 5 in accordance with the type of the image. Moreover, learning is carried out by selecting the learning data in accordance with the type of the image data. The type of the image can be determined according to the following information or method.

(1) A method for performing simple image recognition to find a section where pixel values are relatively close to each other.

(2) Sampling conditions.

(3) Slices or sections (an axial section, a coronal section, and a sagittal section).

(4) Scanning object that can be obtained through a hospital filing system.

(5) Features (e.g., total area, peak position, shape, and height) of the histogram.

Although selection of the image for learning has not been described in the above-mentioned embodiments, for example, ten images (for each image type) are stored as the learning images for each device, and the neural network is caused to learn based on these images. In this way, the neural network (weight coefficient) suitable for the device is formed. If the entire learning images obtained by adding new learning images or updating old images to new images are used again to carry out the learning process, the system can also cope with a change in the device. Alternatively, instead of setting the weight coefficient for each device and for each image type, the weight coefficient may also be set depending on the device, a user (doctor), and the type of image.

In the embodiments described above, the evaluation feature data to be input to the neural network are linearly normalized values. However, normalization may also be carried out using a sine function or a logarithmic transformation.

Even if the number of evaluation features coincides with the number of units in the input layer of the neural network in the above-mentioned embodiments, a plurality of units may be assigned to one evaluation feature, and a value obtained by linear normalization of the evaluation feature and a value obtained by normalization using a sine function or a logarithmic transformation are entered as respective input values.

In the above embodiments, the number of output units is one, and this output unit represents the image quality including both the brightness and the contrast. However, two output units may be used in accordance with the brightness and the contrast such that the brightness is represented as "too bright - appropriate - too dark" (0 to 1). In this case, the image quality is calculated as a weight sum of the two factors from the two output units. Teacher data for brightness is set as WWS = WWG in Fig. 15, and teacher data for contrast is set as WLS = WLG therein.

Furthermore, in the above embodiments, learning and execution are carried out by the same neural network. However, learning may be carried out by another neural network, which may be included in a work station, for example, while execution is carried out by another simple neural network. In addition, from the weight coefficients of the neural network that has completed its learning, only the weights having large values are extracted and calculated only by the software of a computer without using the neural network.

Claims (17)

1. Digital image display device for displaying digital image data by converting picture element values of the digital image data into brightness of a display, with a device (2) for calculating a frequency distribution of the picture element values of the digital image data, marked by means (2) for calculating a plurality of image qualities indicating the clearness of a display image when the digital image data is to be displayed in accordance with a plurality of conversion modes for converting the picture element values into the brightness, and means (2) for selecting a conversion mode indicating a clear image quality for converting the pixel values into the brightness, the selection being made on the basis of the plurality of image qualities. 2. Apparatus according to claim 1, characterized in that the selecting means comprises means for selecting a conversion mode indicating a maximum image quality for converting the picture element values into the brightness. 3. Device according to claim 1, characterized in that the conversion mode is a display window on a display unit, and that the device (2) for calculating the image quality comprises: a first device for calculating the image qualities in the display window set as a candidate and in a display window shifted from the candidate by a predetermined amount, a second device which is designed, when the image quality in the display window different from the candidate is maximum, to input the display window having the maximum image quality as a new candidate, and a third means for reducing the predetermined size used by the first means when the image quality of the candidate is maximum. 4. Device according to claim 1, characterized in that the device for calculating the image quality comprises: first means for calculating first data indicating the brightness of the display image, second data indicating the contrast of the display image, and third data indicating the brightness balance, the calculations being performed depending on the frequency distribution of the picture element values, and second means for calculating the image quality based on the first to third data. 5. An apparatus according to claim 4, characterized in that said second means includes means for adding data obtained by converting said first to third data with a non-linear function and weighted by a predetermined value for calculating image quality. 6. Apparatus according to claim 4, characterized in that the second device includes a neural network (5) for inputting normalized data from the first to third data and for outputting the image quality. 7. An apparatus according to claim 4, characterized in that said first data includes brightness data for a pixel value having the highest frequency and brightness data for a pixel value at a boundary between a background and an image, said data being included in a histogram indicating frequency distribution, said second data includes range data for a portion having average brightness within a display brightness range and area data for a portion having the maximum brightness within the range, and said third data includes data indicating a relationship between an area of a area of a portion having a brightness higher than the mean brightness and an area of a portion having a brightness lower than this. 8. An apparatus according to claim 1, characterized in that the image quality calculating means (2) comprises means for calculating the image quality indicative of the clarity of each region of an image to be observed, and that the selecting means comprises means for selecting a conversion mode for each section. 9. Device according to claim B, characterized in that each section is a section which corresponds to a respective peak frequency of the frequency distribution. 10. Device according to one of the preceding claims, characterized in that the conversion mode is defined by a window width and a window level. 11. Digital image display device for converting picture element values of digital image data into brightness values, such that only picture element values within a predetermined range are converted into continuously changing brightness values, with a device for calculating a frequency distribution of the pixel values of the digital image data, marked by a neural network device (5) for receiving a plurality of specific parameters in the frequency distribution and for outputting an image quality indicating the clarity of the image displayed with the brightness value, and means (2) for determining an optimum value of the predetermined range in dependence on the output signal from the neural network means (5) obtained when the predetermined range is changed. 12. Device according to claim 11, characterized in that the device for determining the optimal value comprises: a device for determining a candidate for the predetermined area, means for calculating the image quality of the candidate and the image quality obtained when the center of the predetermined area is changed or when the predetermined area is changed, and means for changing the candidate so that the image quality of the candidate is maximized, and for repeatedly calculating and comparing the image quality of the candidate and the image quality obtained when the center of the predetermined area is changed or when the predetermined area is changed. 13. An apparatus according to claim 11 or 12, characterized in that the neural network means comprises at least two neural networks (5a, 5b) each of which includes at least three layers and precisely learns different data, and in that the optimum value determining means (2) determines an optimum value by using the neural networks (5a, 5b) sequentially starting from a neural network that learns data with low precision, so that an output signal generated by each neural network (5a, 5b) is used as a candidate in processing performed by the next neural network. 14. An apparatus according to claim 11, 12 or 13, characterized in that the neural network device (5) receives first data indicating the brightness of a display image, second data indicating the contrast of the display image, and third data indicating the brightness balance based on the frequency distribution. 15. Apparatus according to claim 14, characterized in that the first data are brightness data of a picture element value having a highest frequency and brightness data of a picture element value located at a boundary between a background and an image on a histogram indicating frequency distribution, the second data includes area data of a region having average brightness within the display brightness range and area data of a region having maximum brightness within the range, and the third data includes data indicating a relationship between an area of a region having higher brightness than the average brightness and an area of a region having lower brightness than the average brightness. 16. An apparatus according to any one of claims 11 to 15, characterized in that the neural network means (5) generates the image quality indicating the clarity of each portion to be observed in an image, and the brightness data of a pixel value having the highest frequency includes brightness data of a pixel value at each peak in the frequency distribution. 17. Device according to claim 1, which serves to determine an optimal window level and an optimal window width, characterized by a first image quality calculation device which indicates, by means of an index, the clarity of an image when the window level and the window width are respectively set to form predetermined candidate values, a second device for calculating the image quality when the window level and/or the window width are shifted by a predetermined value from the candidate values, a third device for comparing the image quality calculated by the first device and the image quality calculated by the second device, and means for changing the candidate values and for reducing the predetermined value depending on the comparison result, for repeatedly operating the first, second and third means such that a maximum image quality is obtained by the first means, and for setting the candidate values which set in the initial setup as the optimal window level and window width.